---
title: An OpenAI Researcher Has Left to Build Models That Turn Thoughts Into Text
description: Naomi Bashkansky has joined Conduit, a startup training models to turn brain activity into text, and it pays people in San Francisco to supply the data.
author: Darie Nani (Editor-in-Chief)
updated: 2026-08-06T21:25:22.713Z
canonical: https://www.sovereignmagazine.com/article/openai-researcher-conduit-thought-to-text
image: https://cdn.nanimediahouse.com/naomi-bashkansky-conduit-113817.webp
categories: Startups
content_type: News
region: United States
publication: Sovereign Magazine
schema_type: Article
---

Naomi Bashkansky resigned from OpenAI on 23 July and started the next day as a founding researcher at Conduit, a startup training models to turn brain activity into text without an implant. She set out the move in a post on X on 5 August and in a longer piece on her own site the day before.

Conduit describes what it is building on its website: “Put on a headset, start typing, and the software completes your sentences from your thoughts.” It says its aim is “high-bandwidth direct brain-to-computer communication”.

## The Company Pays People to Have Their Brain Activity Recorded

Conduit is recruiting paid research participants at its San Francisco office. They speak and type responses to on-screen prompts while wearing a headset. The published terms are up to $50 for a completed two-hour session and up to $100 for four hours, paid by check, with a maximum of ten two-hour slots per person and no more than two in any 24 hours.

Those sessions produce the training data. Bashkansky [wrote in her own account of the move](https://naomibashkansky.com/blog/telepathy/) that the input is a person's brain activity and the target output is what they were doing at the time, such as the text they wrote, with the model trained to predict something semantically close to it. Collecting enough of that data is the constraint. “It's hard to collect enough data using invasive methods. Few people want a chip in their brain,” she wrote, adding that non-invasive hardware is getting better and cheaper. She did not describe Conduit's own hardware.

## Accuracy Rises With the Hours of Data Collected

Bashkansky said Conduit measures progress by how closely its models' predictions match the target, and that the similarity rises in a straight line against the logarithm of the hours of data collected.

> "We're in the GPT-2 era."
> — Naomi Bashkansky, founding researcher at Conduit

Decoding does not have to be perfect to be useful, she said. She compared a person's thoughts to a noisy GPS signal, not enough on its own to fix a location but accurate when combined with a map and a route. The map and the route, in her comparison, are the language model and its context.

## She Predicts Thought Will Become the Main Way People Talk to AI

Bashkansky expects that within a couple of years people will talk to AI systems mainly through their thoughts rather than by typing. She sketches 2027, 2030 and 2035 in her post. In the 2027 scene she wears a band on her head paired to a laptop over Bluetooth, and a coding agent receives her decoded thoughts in ten-second chunks with auto-send on. By 2030 she imagines AI companies training their models to read Conduit's representations directly, and Conduit starting work on reading through implants and on writing to the brain. These are her predictions rather than company commitments.

## She Left OpenAI for Wider Scope

Bashkansky spent a year and a half as a researcher at OpenAI. There, she said, “you can only work on a narrow slice of The Problem, and even on that narrow slice you're constrained by the existing architecture”, while at Conduit the work is “all greenfield”. She named the co-founders as Rio Popper and Clem von Stengel, and gave three reasons for joining: the vision, the people and the problems. She said she expects the company to be worth more than $1 trillion if it becomes what she calls the general read and write company, which she states as her own view.

She grew up in Washington State and played competitive chess from the age of five until she was fifteen, stopping after becoming a Woman International Master. She said she first heard about the possibility of smarter-than-human AI when she was twelve.

## FAQ

**Q: What does thought-to-text mean?**
Software that turns recorded brain activity into written words, so what a person is thinking reaches a computer without them typing it.

**Q: What does non-invasive mean here?**
The brain activity is recorded from outside the skull, with a headset, rather than through a device implanted in the brain.

**Q: Who can take part in Conduit's sessions?**
Conduit recruits paid participants for sessions at its San Francisco office, where they speak and type responses to on-screen prompts while wearing a headset.